1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation analyzes and transforms the induction variables (and
11 // computations derived from them) into forms suitable for efficient execution
14 // This pass performs a strength reduction on array references inside loops that
15 // have as one or more of their components the loop induction variable, it
16 // rewrites expressions to take advantage of scaled-index addressing modes
17 // available on the target, and it performs a variety of other optimizations
18 // related to loop induction variables.
20 // Terminology note: this code has a lot of handling for "post-increment" or
21 // "post-inc" users. This is not talking about post-increment addressing modes;
22 // it is instead talking about code like this:
24 // %i = phi [ 0, %entry ], [ %i.next, %latch ]
26 // %i.next = add %i, 1
27 // %c = icmp eq %i.next, %n
29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however
30 // it's useful to think about these as the same register, with some uses using
31 // the value of the register before the add and some using // it after. In this
32 // example, the icmp is a post-increment user, since it uses %i.next, which is
33 // the value of the induction variable after the increment. The other common
34 // case of post-increment users is users outside the loop.
36 // TODO: More sophistication in the way Formulae are generated and filtered.
38 // TODO: Handle multiple loops at a time.
40 // TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr
41 // instead of a GlobalValue?
43 // TODO: When truncation is free, truncate ICmp users' operands to make it a
44 // smaller encoding (on x86 at least).
46 // TODO: When a negated register is used by an add (such as in a list of
47 // multiple base registers, or as the increment expression in an addrec),
48 // we may not actually need both reg and (-1 * reg) in registers; the
49 // negation can be implemented by using a sub instead of an add. The
50 // lack of support for taking this into consideration when making
51 // register pressure decisions is partly worked around by the "Special"
54 //===----------------------------------------------------------------------===//
56 #define DEBUG_TYPE "loop-reduce"
57 #include "llvm/Transforms/Scalar.h"
58 #include "llvm/Constants.h"
59 #include "llvm/Instructions.h"
60 #include "llvm/IntrinsicInst.h"
61 #include "llvm/DerivedTypes.h"
62 #include "llvm/Analysis/IVUsers.h"
63 #include "llvm/Analysis/Dominators.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolutionExpander.h"
66 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
67 #include "llvm/Transforms/Utils/Local.h"
68 #include "llvm/ADT/SmallBitVector.h"
69 #include "llvm/ADT/SetVector.h"
70 #include "llvm/ADT/DenseSet.h"
71 #include "llvm/Support/Debug.h"
72 #include "llvm/Support/ValueHandle.h"
73 #include "llvm/Support/raw_ostream.h"
74 #include "llvm/Target/TargetLowering.h"
80 /// RegSortData - This class holds data which is used to order reuse candidates.
83 /// UsedByIndices - This represents the set of LSRUse indices which reference
84 /// a particular register.
85 SmallBitVector UsedByIndices;
89 void print(raw_ostream &OS) const;
95 void RegSortData::print(raw_ostream &OS) const {
96 OS << "[NumUses=" << UsedByIndices.count() << ']';
99 void RegSortData::dump() const {
100 print(errs()); errs() << '\n';
105 /// RegUseTracker - Map register candidates to information about how they are
107 class RegUseTracker {
108 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy;
111 SmallVector<const SCEV *, 16> RegSequence;
114 void CountRegister(const SCEV *Reg, size_t LUIdx);
116 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
118 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const;
122 typedef SmallVectorImpl<const SCEV *>::iterator iterator;
123 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator;
124 iterator begin() { return RegSequence.begin(); }
125 iterator end() { return RegSequence.end(); }
126 const_iterator begin() const { return RegSequence.begin(); }
127 const_iterator end() const { return RegSequence.end(); }
133 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) {
134 std::pair<RegUsesTy::iterator, bool> Pair =
135 RegUses.insert(std::make_pair(Reg, RegSortData()));
136 RegSortData &RSD = Pair.first->second;
138 RegSequence.push_back(Reg);
139 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1));
140 RSD.UsedByIndices.set(LUIdx);
144 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const {
145 if (!RegUses.count(Reg)) return false;
146 const SmallBitVector &UsedByIndices =
147 RegUses.find(Reg)->second.UsedByIndices;
148 int i = UsedByIndices.find_first();
149 if (i == -1) return false;
150 if ((size_t)i != LUIdx) return true;
151 return UsedByIndices.find_next(i) != -1;
154 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const {
155 RegUsesTy::const_iterator I = RegUses.find(Reg);
156 assert(I != RegUses.end() && "Unknown register!");
157 return I->second.UsedByIndices;
160 void RegUseTracker::clear() {
167 /// Formula - This class holds information that describes a formula for
168 /// computing satisfying a use. It may include broken-out immediates and scaled
171 /// AM - This is used to represent complex addressing, as well as other kinds
172 /// of interesting uses.
173 TargetLowering::AddrMode AM;
175 /// BaseRegs - The list of "base" registers for this use. When this is
176 /// non-empty, AM.HasBaseReg should be set to true.
177 SmallVector<const SCEV *, 2> BaseRegs;
179 /// ScaledReg - The 'scaled' register for this use. This should be non-null
180 /// when AM.Scale is not zero.
181 const SCEV *ScaledReg;
183 Formula() : ScaledReg(0) {}
185 void InitialMatch(const SCEV *S, Loop *L,
186 ScalarEvolution &SE, DominatorTree &DT);
188 unsigned getNumRegs() const;
189 const Type *getType() const;
191 bool referencesReg(const SCEV *S) const;
192 bool hasRegsUsedByUsesOtherThan(size_t LUIdx,
193 const RegUseTracker &RegUses) const;
195 void print(raw_ostream &OS) const;
201 /// DoInitialMatch - Recursion helper for InitialMatch.
202 static void DoInitialMatch(const SCEV *S, Loop *L,
203 SmallVectorImpl<const SCEV *> &Good,
204 SmallVectorImpl<const SCEV *> &Bad,
205 ScalarEvolution &SE, DominatorTree &DT) {
206 // Collect expressions which properly dominate the loop header.
207 if (S->properlyDominates(L->getHeader(), &DT)) {
212 // Look at add operands.
213 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
214 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
216 DoInitialMatch(*I, L, Good, Bad, SE, DT);
220 // Look at addrec operands.
221 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
222 if (!AR->getStart()->isZero()) {
223 DoInitialMatch(AR->getStart(), L, Good, Bad, SE, DT);
224 DoInitialMatch(SE.getAddRecExpr(SE.getIntegerSCEV(0, AR->getType()),
225 AR->getStepRecurrence(SE),
227 L, Good, Bad, SE, DT);
231 // Handle a multiplication by -1 (negation) if it didn't fold.
232 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S))
233 if (Mul->getOperand(0)->isAllOnesValue()) {
234 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end());
235 const SCEV *NewMul = SE.getMulExpr(Ops);
237 SmallVector<const SCEV *, 4> MyGood;
238 SmallVector<const SCEV *, 4> MyBad;
239 DoInitialMatch(NewMul, L, MyGood, MyBad, SE, DT);
240 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue(
241 SE.getEffectiveSCEVType(NewMul->getType())));
242 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(),
243 E = MyGood.end(); I != E; ++I)
244 Good.push_back(SE.getMulExpr(NegOne, *I));
245 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(),
246 E = MyBad.end(); I != E; ++I)
247 Bad.push_back(SE.getMulExpr(NegOne, *I));
251 // Ok, we can't do anything interesting. Just stuff the whole thing into a
252 // register and hope for the best.
256 /// InitialMatch - Incorporate loop-variant parts of S into this Formula,
257 /// attempting to keep all loop-invariant and loop-computable values in a
258 /// single base register.
259 void Formula::InitialMatch(const SCEV *S, Loop *L,
260 ScalarEvolution &SE, DominatorTree &DT) {
261 SmallVector<const SCEV *, 4> Good;
262 SmallVector<const SCEV *, 4> Bad;
263 DoInitialMatch(S, L, Good, Bad, SE, DT);
265 const SCEV *Sum = SE.getAddExpr(Good);
267 BaseRegs.push_back(Sum);
268 AM.HasBaseReg = true;
271 const SCEV *Sum = SE.getAddExpr(Bad);
273 BaseRegs.push_back(Sum);
274 AM.HasBaseReg = true;
278 /// getNumRegs - Return the total number of register operands used by this
279 /// formula. This does not include register uses implied by non-constant
281 unsigned Formula::getNumRegs() const {
282 return !!ScaledReg + BaseRegs.size();
285 /// getType - Return the type of this formula, if it has one, or null
286 /// otherwise. This type is meaningless except for the bit size.
287 const Type *Formula::getType() const {
288 return !BaseRegs.empty() ? BaseRegs.front()->getType() :
289 ScaledReg ? ScaledReg->getType() :
290 AM.BaseGV ? AM.BaseGV->getType() :
294 /// referencesReg - Test if this formula references the given register.
295 bool Formula::referencesReg(const SCEV *S) const {
296 return S == ScaledReg ||
297 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end();
300 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers
301 /// which are used by uses other than the use with the given index.
302 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx,
303 const RegUseTracker &RegUses) const {
305 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx))
307 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
308 E = BaseRegs.end(); I != E; ++I)
309 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx))
314 void Formula::print(raw_ostream &OS) const {
317 if (!First) OS << " + "; else First = false;
318 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false);
320 if (AM.BaseOffs != 0) {
321 if (!First) OS << " + "; else First = false;
324 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(),
325 E = BaseRegs.end(); I != E; ++I) {
326 if (!First) OS << " + "; else First = false;
327 OS << "reg(" << **I << ')';
330 if (!First) OS << " + "; else First = false;
331 OS << AM.Scale << "*reg(";
340 void Formula::dump() const {
341 print(errs()); errs() << '\n';
344 /// isAddRecSExtable - Return true if the given addrec can be sign-extended
345 /// without changing its value.
346 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) {
348 IntegerType::get(SE.getContext(),
349 SE.getTypeSizeInBits(AR->getType()) + 1);
350 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy));
353 /// isAddSExtable - Return true if the given add can be sign-extended
354 /// without changing its value.
355 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) {
357 IntegerType::get(SE.getContext(),
358 SE.getTypeSizeInBits(A->getType()) + 1);
359 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy));
362 /// isMulSExtable - Return true if the given add can be sign-extended
363 /// without changing its value.
364 static bool isMulSExtable(const SCEVMulExpr *A, ScalarEvolution &SE) {
366 IntegerType::get(SE.getContext(),
367 SE.getTypeSizeInBits(A->getType()) + 1);
368 return isa<SCEVMulExpr>(SE.getSignExtendExpr(A, WideTy));
371 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined
372 /// and if the remainder is known to be zero, or null otherwise. If
373 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified
374 /// to Y, ignoring that the multiplication may overflow, which is useful when
375 /// the result will be used in a context where the most significant bits are
377 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS,
379 bool IgnoreSignificantBits = false) {
380 // Handle the trivial case, which works for any SCEV type.
382 return SE.getIntegerSCEV(1, LHS->getType());
384 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do some
386 if (RHS->isAllOnesValue())
387 return SE.getMulExpr(LHS, RHS);
389 // Check for a division of a constant by a constant.
390 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) {
391 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS);
394 if (C->getValue()->getValue().srem(RC->getValue()->getValue()) != 0)
396 return SE.getConstant(C->getValue()->getValue()
397 .sdiv(RC->getValue()->getValue()));
400 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow.
401 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) {
402 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) {
403 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE,
404 IgnoreSignificantBits);
405 if (!Start) return 0;
406 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE,
407 IgnoreSignificantBits);
409 return SE.getAddRecExpr(Start, Step, AR->getLoop());
413 // Distribute the sdiv over add operands, if the add doesn't overflow.
414 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) {
415 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) {
416 SmallVector<const SCEV *, 8> Ops;
417 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
419 const SCEV *Op = getExactSDiv(*I, RHS, SE,
420 IgnoreSignificantBits);
424 return SE.getAddExpr(Ops);
428 // Check for a multiply operand that we can pull RHS out of.
429 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS))
430 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) {
431 SmallVector<const SCEV *, 4> Ops;
433 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end();
436 if (const SCEV *Q = getExactSDiv(*I, RHS, SE,
437 IgnoreSignificantBits)) {
444 return Found ? SE.getMulExpr(Ops) : 0;
447 // Otherwise we don't know.
451 /// ExtractImmediate - If S involves the addition of a constant integer value,
452 /// return that integer value, and mutate S to point to a new SCEV with that
454 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) {
455 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
456 if (C->getValue()->getValue().getMinSignedBits() <= 64) {
457 S = SE.getIntegerSCEV(0, C->getType());
458 return C->getValue()->getSExtValue();
460 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
461 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
462 int64_t Result = ExtractImmediate(NewOps.front(), SE);
463 S = SE.getAddExpr(NewOps);
465 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
466 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
467 int64_t Result = ExtractImmediate(NewOps.front(), SE);
468 S = SE.getAddRecExpr(NewOps, AR->getLoop());
474 /// ExtractSymbol - If S involves the addition of a GlobalValue address,
475 /// return that symbol, and mutate S to point to a new SCEV with that
477 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) {
478 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
479 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) {
480 S = SE.getIntegerSCEV(0, GV->getType());
483 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
484 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end());
485 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE);
486 S = SE.getAddExpr(NewOps);
488 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
489 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end());
490 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE);
491 S = SE.getAddRecExpr(NewOps, AR->getLoop());
497 /// isAddressUse - Returns true if the specified instruction is using the
498 /// specified value as an address.
499 static bool isAddressUse(Instruction *Inst, Value *OperandVal) {
500 bool isAddress = isa<LoadInst>(Inst);
501 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
502 if (SI->getOperand(1) == OperandVal)
504 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
505 // Addressing modes can also be folded into prefetches and a variety
507 switch (II->getIntrinsicID()) {
509 case Intrinsic::prefetch:
510 case Intrinsic::x86_sse2_loadu_dq:
511 case Intrinsic::x86_sse2_loadu_pd:
512 case Intrinsic::x86_sse_loadu_ps:
513 case Intrinsic::x86_sse_storeu_ps:
514 case Intrinsic::x86_sse2_storeu_pd:
515 case Intrinsic::x86_sse2_storeu_dq:
516 case Intrinsic::x86_sse2_storel_dq:
517 if (II->getOperand(1) == OperandVal)
525 /// getAccessType - Return the type of the memory being accessed.
526 static const Type *getAccessType(const Instruction *Inst) {
527 const Type *AccessTy = Inst->getType();
528 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst))
529 AccessTy = SI->getOperand(0)->getType();
530 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
531 // Addressing modes can also be folded into prefetches and a variety
533 switch (II->getIntrinsicID()) {
535 case Intrinsic::x86_sse_storeu_ps:
536 case Intrinsic::x86_sse2_storeu_pd:
537 case Intrinsic::x86_sse2_storeu_dq:
538 case Intrinsic::x86_sse2_storel_dq:
539 AccessTy = II->getOperand(1)->getType();
544 // All pointers have the same requirements, so canonicalize them to an
545 // arbitrary pointer type to minimize variation.
546 if (const PointerType *PTy = dyn_cast<PointerType>(AccessTy))
547 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1),
548 PTy->getAddressSpace());
553 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
554 /// specified set are trivially dead, delete them and see if this makes any of
555 /// their operands subsequently dead.
557 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) {
558 bool Changed = false;
560 while (!DeadInsts.empty()) {
561 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
563 if (I == 0 || !isInstructionTriviallyDead(I))
566 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
567 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
570 DeadInsts.push_back(U);
573 I->eraseFromParent();
582 /// Cost - This class is used to measure and compare candidate formulae.
584 /// TODO: Some of these could be merged. Also, a lexical ordering
585 /// isn't always optimal.
589 unsigned NumBaseAdds;
595 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0),
598 unsigned getNumRegs() const { return NumRegs; }
600 bool operator<(const Cost &Other) const;
604 void RateFormula(const Formula &F,
605 SmallPtrSet<const SCEV *, 16> &Regs,
606 const DenseSet<const SCEV *> &VisitedRegs,
608 const SmallVectorImpl<int64_t> &Offsets,
609 ScalarEvolution &SE, DominatorTree &DT);
611 void print(raw_ostream &OS) const;
615 void RateRegister(const SCEV *Reg,
616 SmallPtrSet<const SCEV *, 16> &Regs,
618 ScalarEvolution &SE, DominatorTree &DT);
619 void RatePrimaryRegister(const SCEV *Reg,
620 SmallPtrSet<const SCEV *, 16> &Regs,
622 ScalarEvolution &SE, DominatorTree &DT);
627 /// RateRegister - Tally up interesting quantities from the given register.
628 void Cost::RateRegister(const SCEV *Reg,
629 SmallPtrSet<const SCEV *, 16> &Regs,
631 ScalarEvolution &SE, DominatorTree &DT) {
632 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) {
633 if (AR->getLoop() == L)
634 AddRecCost += 1; /// TODO: This should be a function of the stride.
636 // If this is an addrec for a loop that's already been visited by LSR,
637 // don't second-guess its addrec phi nodes. LSR isn't currently smart
638 // enough to reason about more than one loop at a time. Consider these
639 // registers free and leave them alone.
640 else if (L->contains(AR->getLoop()) ||
641 (!AR->getLoop()->contains(L) &&
642 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) {
643 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin();
644 PHINode *PN = dyn_cast<PHINode>(I); ++I)
645 if (SE.isSCEVable(PN->getType()) &&
646 (SE.getEffectiveSCEVType(PN->getType()) ==
647 SE.getEffectiveSCEVType(AR->getType())) &&
648 SE.getSCEV(PN) == AR)
651 // If this isn't one of the addrecs that the loop already has, it
652 // would require a costly new phi and add. TODO: This isn't
653 // precisely modeled right now.
655 if (!Regs.count(AR->getStart()))
656 RateRegister(AR->getStart(), Regs, L, SE, DT);
659 // Add the step value register, if it needs one.
660 // TODO: The non-affine case isn't precisely modeled here.
661 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1)))
662 if (!Regs.count(AR->getStart()))
663 RateRegister(AR->getOperand(1), Regs, L, SE, DT);
667 // Rough heuristic; favor registers which don't require extra setup
668 // instructions in the preheader.
669 if (!isa<SCEVUnknown>(Reg) &&
670 !isa<SCEVConstant>(Reg) &&
671 !(isa<SCEVAddRecExpr>(Reg) &&
672 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) ||
673 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart()))))
677 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it
679 void Cost::RatePrimaryRegister(const SCEV *Reg,
680 SmallPtrSet<const SCEV *, 16> &Regs,
682 ScalarEvolution &SE, DominatorTree &DT) {
683 if (Regs.insert(Reg))
684 RateRegister(Reg, Regs, L, SE, DT);
687 void Cost::RateFormula(const Formula &F,
688 SmallPtrSet<const SCEV *, 16> &Regs,
689 const DenseSet<const SCEV *> &VisitedRegs,
691 const SmallVectorImpl<int64_t> &Offsets,
692 ScalarEvolution &SE, DominatorTree &DT) {
693 // Tally up the registers.
694 if (const SCEV *ScaledReg = F.ScaledReg) {
695 if (VisitedRegs.count(ScaledReg)) {
699 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT);
701 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
702 E = F.BaseRegs.end(); I != E; ++I) {
703 const SCEV *BaseReg = *I;
704 if (VisitedRegs.count(BaseReg)) {
708 RatePrimaryRegister(BaseReg, Regs, L, SE, DT);
710 NumIVMuls += isa<SCEVMulExpr>(BaseReg) &&
711 BaseReg->hasComputableLoopEvolution(L);
714 if (F.BaseRegs.size() > 1)
715 NumBaseAdds += F.BaseRegs.size() - 1;
717 // Tally up the non-zero immediates.
718 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
719 E = Offsets.end(); I != E; ++I) {
720 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs;
722 ImmCost += 64; // Handle symbolic values conservatively.
723 // TODO: This should probably be the pointer size.
724 else if (Offset != 0)
725 ImmCost += APInt(64, Offset, true).getMinSignedBits();
729 /// Loose - Set this cost to a loosing value.
739 /// operator< - Choose the lower cost.
740 bool Cost::operator<(const Cost &Other) const {
741 if (NumRegs != Other.NumRegs)
742 return NumRegs < Other.NumRegs;
743 if (AddRecCost != Other.AddRecCost)
744 return AddRecCost < Other.AddRecCost;
745 if (NumIVMuls != Other.NumIVMuls)
746 return NumIVMuls < Other.NumIVMuls;
747 if (NumBaseAdds != Other.NumBaseAdds)
748 return NumBaseAdds < Other.NumBaseAdds;
749 if (ImmCost != Other.ImmCost)
750 return ImmCost < Other.ImmCost;
751 if (SetupCost != Other.SetupCost)
752 return SetupCost < Other.SetupCost;
756 void Cost::print(raw_ostream &OS) const {
757 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s");
759 OS << ", with addrec cost " << AddRecCost;
761 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s");
762 if (NumBaseAdds != 0)
763 OS << ", plus " << NumBaseAdds << " base add"
764 << (NumBaseAdds == 1 ? "" : "s");
766 OS << ", plus " << ImmCost << " imm cost";
768 OS << ", plus " << SetupCost << " setup cost";
771 void Cost::dump() const {
772 print(errs()); errs() << '\n';
777 /// LSRFixup - An operand value in an instruction which is to be replaced
778 /// with some equivalent, possibly strength-reduced, replacement.
780 /// UserInst - The instruction which will be updated.
781 Instruction *UserInst;
783 /// OperandValToReplace - The operand of the instruction which will
784 /// be replaced. The operand may be used more than once; every instance
785 /// will be replaced.
786 Value *OperandValToReplace;
788 /// PostIncLoops - If this user is to use the post-incremented value of an
789 /// induction variable, this variable is non-null and holds the loop
790 /// associated with the induction variable.
791 PostIncLoopSet PostIncLoops;
793 /// LUIdx - The index of the LSRUse describing the expression which
794 /// this fixup needs, minus an offset (below).
797 /// Offset - A constant offset to be added to the LSRUse expression.
798 /// This allows multiple fixups to share the same LSRUse with different
799 /// offsets, for example in an unrolled loop.
802 bool isUseFullyOutsideLoop(const Loop *L) const;
806 void print(raw_ostream &OS) const;
813 : UserInst(0), OperandValToReplace(0),
814 LUIdx(~size_t(0)), Offset(0) {}
816 /// isUseFullyOutsideLoop - Test whether this fixup always uses its
817 /// value outside of the given loop.
818 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
819 // PHI nodes use their value in their incoming blocks.
820 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
821 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
822 if (PN->getIncomingValue(i) == OperandValToReplace &&
823 L->contains(PN->getIncomingBlock(i)))
828 return !L->contains(UserInst);
831 void LSRFixup::print(raw_ostream &OS) const {
833 // Store is common and interesting enough to be worth special-casing.
834 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) {
836 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false);
837 } else if (UserInst->getType()->isVoidTy())
838 OS << UserInst->getOpcodeName();
840 WriteAsOperand(OS, UserInst, /*PrintType=*/false);
842 OS << ", OperandValToReplace=";
843 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
845 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
846 E = PostIncLoops.end(); I != E; ++I) {
847 OS << ", PostIncLoop=";
848 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
851 if (LUIdx != ~size_t(0))
852 OS << ", LUIdx=" << LUIdx;
855 OS << ", Offset=" << Offset;
858 void LSRFixup::dump() const {
859 print(errs()); errs() << '\n';
864 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding
865 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*.
866 struct UniquifierDenseMapInfo {
867 static SmallVector<const SCEV *, 2> getEmptyKey() {
868 SmallVector<const SCEV *, 2> V;
869 V.push_back(reinterpret_cast<const SCEV *>(-1));
873 static SmallVector<const SCEV *, 2> getTombstoneKey() {
874 SmallVector<const SCEV *, 2> V;
875 V.push_back(reinterpret_cast<const SCEV *>(-2));
879 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) {
881 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(),
882 E = V.end(); I != E; ++I)
883 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I);
887 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS,
888 const SmallVector<const SCEV *, 2> &RHS) {
893 /// LSRUse - This class holds the state that LSR keeps for each use in
894 /// IVUsers, as well as uses invented by LSR itself. It includes information
895 /// about what kinds of things can be folded into the user, information about
896 /// the user itself, and information about how the use may be satisfied.
897 /// TODO: Represent multiple users of the same expression in common?
899 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier;
902 /// KindType - An enum for a kind of use, indicating what types of
903 /// scaled and immediate operands it might support.
905 Basic, ///< A normal use, with no folding.
906 Special, ///< A special case of basic, allowing -1 scales.
907 Address, ///< An address use; folding according to TargetLowering
908 ICmpZero ///< An equality icmp with both operands folded into one.
909 // TODO: Add a generic icmp too?
913 const Type *AccessTy;
915 SmallVector<int64_t, 8> Offsets;
919 /// AllFixupsOutsideLoop - This records whether all of the fixups using this
920 /// LSRUse are outside of the loop, in which case some special-case heuristics
922 bool AllFixupsOutsideLoop;
924 /// Formulae - A list of ways to build a value that can satisfy this user.
925 /// After the list is populated, one of these is selected heuristically and
926 /// used to formulate a replacement for OperandValToReplace in UserInst.
927 SmallVector<Formula, 12> Formulae;
929 /// Regs - The set of register candidates used by all formulae in this LSRUse.
930 SmallPtrSet<const SCEV *, 4> Regs;
932 LSRUse(KindType K, const Type *T) : Kind(K), AccessTy(T),
933 MinOffset(INT64_MAX),
934 MaxOffset(INT64_MIN),
935 AllFixupsOutsideLoop(true) {}
937 bool InsertFormula(const Formula &F);
941 void print(raw_ostream &OS) const;
945 /// InsertFormula - If the given formula has not yet been inserted, add it to
946 /// the list, and return true. Return false otherwise.
947 bool LSRUse::InsertFormula(const Formula &F) {
948 SmallVector<const SCEV *, 2> Key = F.BaseRegs;
949 if (F.ScaledReg) Key.push_back(F.ScaledReg);
950 // Unstable sort by host order ok, because this is only used for uniquifying.
951 std::sort(Key.begin(), Key.end());
953 if (!Uniquifier.insert(Key).second)
956 // Using a register to hold the value of 0 is not profitable.
957 assert((!F.ScaledReg || !F.ScaledReg->isZero()) &&
958 "Zero allocated in a scaled register!");
960 for (SmallVectorImpl<const SCEV *>::const_iterator I =
961 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I)
962 assert(!(*I)->isZero() && "Zero allocated in a base register!");
965 // Add the formula to the list.
966 Formulae.push_back(F);
968 // Record registers now being used by this use.
969 if (F.ScaledReg) Regs.insert(F.ScaledReg);
970 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
975 void LSRUse::print(raw_ostream &OS) const {
976 OS << "LSR Use: Kind=";
978 case Basic: OS << "Basic"; break;
979 case Special: OS << "Special"; break;
980 case ICmpZero: OS << "ICmpZero"; break;
983 if (AccessTy->isPointerTy())
984 OS << "pointer"; // the full pointer type could be really verbose
990 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(),
991 E = Offsets.end(); I != E; ++I) {
998 if (AllFixupsOutsideLoop)
999 OS << ", all-fixups-outside-loop";
1002 void LSRUse::dump() const {
1003 print(errs()); errs() << '\n';
1006 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can
1007 /// be completely folded into the user instruction at isel time. This includes
1008 /// address-mode folding and special icmp tricks.
1009 static bool isLegalUse(const TargetLowering::AddrMode &AM,
1010 LSRUse::KindType Kind, const Type *AccessTy,
1011 const TargetLowering *TLI) {
1013 case LSRUse::Address:
1014 // If we have low-level target information, ask the target if it can
1015 // completely fold this address.
1016 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
1018 // Otherwise, just guess that reg+reg addressing is legal.
1019 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
1021 case LSRUse::ICmpZero:
1022 // There's not even a target hook for querying whether it would be legal to
1023 // fold a GV into an ICmp.
1027 // ICmp only has two operands; don't allow more than two non-trivial parts.
1028 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0)
1031 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by
1032 // putting the scaled register in the other operand of the icmp.
1033 if (AM.Scale != 0 && AM.Scale != -1)
1036 // If we have low-level target information, ask the target if it can fold an
1037 // integer immediate on an icmp.
1038 if (AM.BaseOffs != 0) {
1039 if (TLI) return TLI->isLegalICmpImmediate(-AM.BaseOffs);
1046 // Only handle single-register values.
1047 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0;
1049 case LSRUse::Special:
1050 // Only handle -1 scales, or no scale.
1051 return AM.Scale == 0 || AM.Scale == -1;
1057 static bool isLegalUse(TargetLowering::AddrMode AM,
1058 int64_t MinOffset, int64_t MaxOffset,
1059 LSRUse::KindType Kind, const Type *AccessTy,
1060 const TargetLowering *TLI) {
1061 // Check for overflow.
1062 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
1065 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
1066 if (isLegalUse(AM, Kind, AccessTy, TLI)) {
1067 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
1068 // Check for overflow.
1069 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
1072 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
1073 return isLegalUse(AM, Kind, AccessTy, TLI);
1078 static bool isAlwaysFoldable(int64_t BaseOffs,
1079 GlobalValue *BaseGV,
1081 LSRUse::KindType Kind, const Type *AccessTy,
1082 const TargetLowering *TLI) {
1083 // Fast-path: zero is always foldable.
1084 if (BaseOffs == 0 && !BaseGV) return true;
1086 // Conservatively, create an address with an immediate and a
1087 // base and a scale.
1088 TargetLowering::AddrMode AM;
1089 AM.BaseOffs = BaseOffs;
1091 AM.HasBaseReg = HasBaseReg;
1092 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1094 return isLegalUse(AM, Kind, AccessTy, TLI);
1097 static bool isAlwaysFoldable(const SCEV *S,
1098 int64_t MinOffset, int64_t MaxOffset,
1100 LSRUse::KindType Kind, const Type *AccessTy,
1101 const TargetLowering *TLI,
1102 ScalarEvolution &SE) {
1103 // Fast-path: zero is always foldable.
1104 if (S->isZero()) return true;
1106 // Conservatively, create an address with an immediate and a
1107 // base and a scale.
1108 int64_t BaseOffs = ExtractImmediate(S, SE);
1109 GlobalValue *BaseGV = ExtractSymbol(S, SE);
1111 // If there's anything else involved, it's not foldable.
1112 if (!S->isZero()) return false;
1114 // Fast-path: zero is always foldable.
1115 if (BaseOffs == 0 && !BaseGV) return true;
1117 // Conservatively, create an address with an immediate and a
1118 // base and a scale.
1119 TargetLowering::AddrMode AM;
1120 AM.BaseOffs = BaseOffs;
1122 AM.HasBaseReg = HasBaseReg;
1123 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
1125 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
1128 /// FormulaSorter - This class implements an ordering for formulae which sorts
1129 /// the by their standalone cost.
1130 class FormulaSorter {
1131 /// These two sets are kept empty, so that we compute standalone costs.
1132 DenseSet<const SCEV *> VisitedRegs;
1133 SmallPtrSet<const SCEV *, 16> Regs;
1136 ScalarEvolution &SE;
1140 FormulaSorter(Loop *l, LSRUse &lu, ScalarEvolution &se, DominatorTree &dt)
1141 : L(l), LU(&lu), SE(se), DT(dt) {}
1143 bool operator()(const Formula &A, const Formula &B) {
1145 CostA.RateFormula(A, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1148 CostB.RateFormula(B, Regs, VisitedRegs, L, LU->Offsets, SE, DT);
1150 return CostA < CostB;
1154 /// LSRInstance - This class holds state for the main loop strength reduction
1158 ScalarEvolution &SE;
1161 const TargetLowering *const TLI;
1165 /// IVIncInsertPos - This is the insert position that the current loop's
1166 /// induction variable increment should be placed. In simple loops, this is
1167 /// the latch block's terminator. But in more complicated cases, this is a
1168 /// position which will dominate all the in-loop post-increment users.
1169 Instruction *IVIncInsertPos;
1171 /// Factors - Interesting factors between use strides.
1172 SmallSetVector<int64_t, 8> Factors;
1174 /// Types - Interesting use types, to facilitate truncation reuse.
1175 SmallSetVector<const Type *, 4> Types;
1177 /// Fixups - The list of operands which are to be replaced.
1178 SmallVector<LSRFixup, 16> Fixups;
1180 /// Uses - The list of interesting uses.
1181 SmallVector<LSRUse, 16> Uses;
1183 /// RegUses - Track which uses use which register candidates.
1184 RegUseTracker RegUses;
1186 void OptimizeShadowIV();
1187 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse);
1188 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse);
1189 bool OptimizeLoopTermCond();
1191 void CollectInterestingTypesAndFactors();
1192 void CollectFixupsAndInitialFormulae();
1194 LSRFixup &getNewFixup() {
1195 Fixups.push_back(LSRFixup());
1196 return Fixups.back();
1199 // Support for sharing of LSRUses between LSRFixups.
1200 typedef DenseMap<const SCEV *, size_t> UseMapTy;
1203 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
1204 LSRUse::KindType Kind, const Type *AccessTy);
1206 std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
1207 LSRUse::KindType Kind,
1208 const Type *AccessTy);
1211 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1212 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
1213 void CountRegisters(const Formula &F, size_t LUIdx);
1214 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F);
1216 void CollectLoopInvariantFixupsAndFormulae();
1218 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base,
1219 unsigned Depth = 0);
1220 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base);
1221 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1222 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base);
1223 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1224 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base);
1225 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base);
1226 void GenerateCrossUseConstantOffsets();
1227 void GenerateAllReuseFormulae();
1229 void FilterOutUndesirableDedicatedRegisters();
1230 void NarrowSearchSpaceUsingHeuristics();
1232 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
1234 SmallVectorImpl<const Formula *> &Workspace,
1235 const Cost &CurCost,
1236 const SmallPtrSet<const SCEV *, 16> &CurRegs,
1237 DenseSet<const SCEV *> &VisitedRegs) const;
1238 void Solve(SmallVectorImpl<const Formula *> &Solution) const;
1240 BasicBlock::iterator
1241 HoistInsertPosition(BasicBlock::iterator IP,
1242 const SmallVectorImpl<Instruction *> &Inputs) const;
1243 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP,
1245 const LSRUse &LU) const;
1247 Value *Expand(const LSRFixup &LF,
1249 BasicBlock::iterator IP,
1250 SCEVExpander &Rewriter,
1251 SmallVectorImpl<WeakVH> &DeadInsts) const;
1252 void RewriteForPHI(PHINode *PN, const LSRFixup &LF,
1254 SCEVExpander &Rewriter,
1255 SmallVectorImpl<WeakVH> &DeadInsts,
1257 void Rewrite(const LSRFixup &LF,
1259 SCEVExpander &Rewriter,
1260 SmallVectorImpl<WeakVH> &DeadInsts,
1262 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
1265 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
1267 bool getChanged() const { return Changed; }
1269 void print_factors_and_types(raw_ostream &OS) const;
1270 void print_fixups(raw_ostream &OS) const;
1271 void print_uses(raw_ostream &OS) const;
1272 void print(raw_ostream &OS) const;
1278 /// OptimizeShadowIV - If IV is used in a int-to-float cast
1279 /// inside the loop then try to eliminate the cast operation.
1280 void LSRInstance::OptimizeShadowIV() {
1281 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1282 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1285 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end();
1286 UI != E; /* empty */) {
1287 IVUsers::const_iterator CandidateUI = UI;
1289 Instruction *ShadowUse = CandidateUI->getUser();
1290 const Type *DestTy = NULL;
1292 /* If shadow use is a int->float cast then insert a second IV
1293 to eliminate this cast.
1295 for (unsigned i = 0; i < n; ++i)
1301 for (unsigned i = 0; i < n; ++i, ++d)
1304 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser()))
1305 DestTy = UCast->getDestTy();
1306 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser()))
1307 DestTy = SCast->getDestTy();
1308 if (!DestTy) continue;
1311 // If target does not support DestTy natively then do not apply
1312 // this transformation.
1313 EVT DVT = TLI->getValueType(DestTy);
1314 if (!TLI->isTypeLegal(DVT)) continue;
1317 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
1319 if (PH->getNumIncomingValues() != 2) continue;
1321 const Type *SrcTy = PH->getType();
1322 int Mantissa = DestTy->getFPMantissaWidth();
1323 if (Mantissa == -1) continue;
1324 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa)
1327 unsigned Entry, Latch;
1328 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) {
1336 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry));
1337 if (!Init) continue;
1338 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue());
1340 BinaryOperator *Incr =
1341 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch));
1342 if (!Incr) continue;
1343 if (Incr->getOpcode() != Instruction::Add
1344 && Incr->getOpcode() != Instruction::Sub)
1347 /* Initialize new IV, double d = 0.0 in above example. */
1348 ConstantInt *C = NULL;
1349 if (Incr->getOperand(0) == PH)
1350 C = dyn_cast<ConstantInt>(Incr->getOperand(1));
1351 else if (Incr->getOperand(1) == PH)
1352 C = dyn_cast<ConstantInt>(Incr->getOperand(0));
1358 // Ignore negative constants, as the code below doesn't handle them
1359 // correctly. TODO: Remove this restriction.
1360 if (!C->getValue().isStrictlyPositive()) continue;
1362 /* Add new PHINode. */
1363 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH);
1365 /* create new increment. '++d' in above example. */
1366 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue());
1367 BinaryOperator *NewIncr =
1368 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ?
1369 Instruction::FAdd : Instruction::FSub,
1370 NewPH, CFP, "IV.S.next.", Incr);
1372 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry));
1373 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch));
1375 /* Remove cast operation */
1376 ShadowUse->replaceAllUsesWith(NewPH);
1377 ShadowUse->eraseFromParent();
1382 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV,
1383 /// set the IV user and stride information and return true, otherwise return
1385 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond,
1386 IVStrideUse *&CondUse) {
1387 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1388 if (UI->getUser() == Cond) {
1389 // NOTE: we could handle setcc instructions with multiple uses here, but
1390 // InstCombine does it as well for simple uses, it's not clear that it
1391 // occurs enough in real life to handle.
1398 /// OptimizeMax - Rewrite the loop's terminating condition if it uses
1399 /// a max computation.
1401 /// This is a narrow solution to a specific, but acute, problem. For loops
1407 /// } while (++i < n);
1409 /// the trip count isn't just 'n', because 'n' might not be positive. And
1410 /// unfortunately this can come up even for loops where the user didn't use
1411 /// a C do-while loop. For example, seemingly well-behaved top-test loops
1412 /// will commonly be lowered like this:
1418 /// } while (++i < n);
1421 /// and then it's possible for subsequent optimization to obscure the if
1422 /// test in such a way that indvars can't find it.
1424 /// When indvars can't find the if test in loops like this, it creates a
1425 /// max expression, which allows it to give the loop a canonical
1426 /// induction variable:
1429 /// max = n < 1 ? 1 : n;
1432 /// } while (++i != max);
1434 /// Canonical induction variables are necessary because the loop passes
1435 /// are designed around them. The most obvious example of this is the
1436 /// LoopInfo analysis, which doesn't remember trip count values. It
1437 /// expects to be able to rediscover the trip count each time it is
1438 /// needed, and it does this using a simple analysis that only succeeds if
1439 /// the loop has a canonical induction variable.
1441 /// However, when it comes time to generate code, the maximum operation
1442 /// can be quite costly, especially if it's inside of an outer loop.
1444 /// This function solves this problem by detecting this type of loop and
1445 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting
1446 /// the instructions for the maximum computation.
1448 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) {
1449 // Check that the loop matches the pattern we're looking for.
1450 if (Cond->getPredicate() != CmpInst::ICMP_EQ &&
1451 Cond->getPredicate() != CmpInst::ICMP_NE)
1454 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1));
1455 if (!Sel || !Sel->hasOneUse()) return Cond;
1457 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L);
1458 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount))
1460 const SCEV *One = SE.getIntegerSCEV(1, BackedgeTakenCount->getType());
1462 // Add one to the backedge-taken count to get the trip count.
1463 const SCEV *IterationCount = SE.getAddExpr(BackedgeTakenCount, One);
1465 // Check for a max calculation that matches the pattern.
1466 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount))
1468 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount);
1469 if (Max != SE.getSCEV(Sel)) return Cond;
1471 // To handle a max with more than two operands, this optimization would
1472 // require additional checking and setup.
1473 if (Max->getNumOperands() != 2)
1476 const SCEV *MaxLHS = Max->getOperand(0);
1477 const SCEV *MaxRHS = Max->getOperand(1);
1478 if (!MaxLHS || MaxLHS != One) return Cond;
1479 // Check the relevant induction variable for conformance to
1481 const SCEV *IV = SE.getSCEV(Cond->getOperand(0));
1482 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV);
1483 if (!AR || !AR->isAffine() ||
1484 AR->getStart() != One ||
1485 AR->getStepRecurrence(SE) != One)
1488 assert(AR->getLoop() == L &&
1489 "Loop condition operand is an addrec in a different loop!");
1491 // Check the right operand of the select, and remember it, as it will
1492 // be used in the new comparison instruction.
1494 if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS)
1495 NewRHS = Sel->getOperand(1);
1496 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS)
1497 NewRHS = Sel->getOperand(2);
1498 if (!NewRHS) return Cond;
1500 // Determine the new comparison opcode. It may be signed or unsigned,
1501 // and the original comparison may be either equality or inequality.
1502 CmpInst::Predicate Pred =
1503 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT;
1504 if (Cond->getPredicate() == CmpInst::ICMP_EQ)
1505 Pred = CmpInst::getInversePredicate(Pred);
1507 // Ok, everything looks ok to change the condition into an SLT or SGE and
1508 // delete the max calculation.
1510 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp");
1512 // Delete the max calculation instructions.
1513 Cond->replaceAllUsesWith(NewCond);
1514 CondUse->setUser(NewCond);
1515 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0));
1516 Cond->eraseFromParent();
1517 Sel->eraseFromParent();
1518 if (Cmp->use_empty())
1519 Cmp->eraseFromParent();
1523 /// OptimizeLoopTermCond - Change loop terminating condition to use the
1524 /// postinc iv when possible.
1526 LSRInstance::OptimizeLoopTermCond() {
1527 SmallPtrSet<Instruction *, 4> PostIncs;
1529 BasicBlock *LatchBlock = L->getLoopLatch();
1530 SmallVector<BasicBlock*, 8> ExitingBlocks;
1531 L->getExitingBlocks(ExitingBlocks);
1533 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
1534 BasicBlock *ExitingBlock = ExitingBlocks[i];
1536 // Get the terminating condition for the loop if possible. If we
1537 // can, we want to change it to use a post-incremented version of its
1538 // induction variable, to allow coalescing the live ranges for the IV into
1539 // one register value.
1541 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
1544 // FIXME: Overly conservative, termination condition could be an 'or' etc..
1545 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition()))
1548 // Search IVUsesByStride to find Cond's IVUse if there is one.
1549 IVStrideUse *CondUse = 0;
1550 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition());
1551 if (!FindIVUserForCond(Cond, CondUse))
1554 // If the trip count is computed in terms of a max (due to ScalarEvolution
1555 // being unable to find a sufficient guard, for example), change the loop
1556 // comparison to use SLT or ULT instead of NE.
1557 // One consequence of doing this now is that it disrupts the count-down
1558 // optimization. That's not always a bad thing though, because in such
1559 // cases it may still be worthwhile to avoid a max.
1560 Cond = OptimizeMax(Cond, CondUse);
1562 // If this exiting block dominates the latch block, it may also use
1563 // the post-inc value if it won't be shared with other uses.
1564 // Check for dominance.
1565 if (!DT.dominates(ExitingBlock, LatchBlock))
1568 // Conservatively avoid trying to use the post-inc value in non-latch
1569 // exits if there may be pre-inc users in intervening blocks.
1570 if (LatchBlock != ExitingBlock)
1571 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI)
1572 // Test if the use is reachable from the exiting block. This dominator
1573 // query is a conservative approximation of reachability.
1574 if (&*UI != CondUse &&
1575 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
1576 // Conservatively assume there may be reuse if the quotient of their
1577 // strides could be a legal scale.
1578 const SCEV *A = CondUse->getStride(L);
1579 const SCEV *B = UI->getStride(L);
1580 if (!A || !B) continue;
1581 if (SE.getTypeSizeInBits(A->getType()) !=
1582 SE.getTypeSizeInBits(B->getType())) {
1583 if (SE.getTypeSizeInBits(A->getType()) >
1584 SE.getTypeSizeInBits(B->getType()))
1585 B = SE.getSignExtendExpr(B, A->getType());
1587 A = SE.getSignExtendExpr(A, B->getType());
1589 if (const SCEVConstant *D =
1590 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) {
1591 // Stride of one or negative one can have reuse with non-addresses.
1592 if (D->getValue()->isOne() ||
1593 D->getValue()->isAllOnesValue())
1594 goto decline_post_inc;
1595 // Avoid weird situations.
1596 if (D->getValue()->getValue().getMinSignedBits() >= 64 ||
1597 D->getValue()->getValue().isMinSignedValue())
1598 goto decline_post_inc;
1599 // Without TLI, assume that any stride might be valid, and so any
1600 // use might be shared.
1602 goto decline_post_inc;
1603 // Check for possible scaled-address reuse.
1604 const Type *AccessTy = getAccessType(UI->getUser());
1605 TargetLowering::AddrMode AM;
1606 AM.Scale = D->getValue()->getSExtValue();
1607 if (TLI->isLegalAddressingMode(AM, AccessTy))
1608 goto decline_post_inc;
1609 AM.Scale = -AM.Scale;
1610 if (TLI->isLegalAddressingMode(AM, AccessTy))
1611 goto decline_post_inc;
1615 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: "
1618 // It's possible for the setcc instruction to be anywhere in the loop, and
1619 // possible for it to have multiple users. If it is not immediately before
1620 // the exiting block branch, move it.
1621 if (&*++BasicBlock::iterator(Cond) != TermBr) {
1622 if (Cond->hasOneUse()) {
1623 Cond->moveBefore(TermBr);
1625 // Clone the terminating condition and insert into the loopend.
1626 ICmpInst *OldCond = Cond;
1627 Cond = cast<ICmpInst>(Cond->clone());
1628 Cond->setName(L->getHeader()->getName() + ".termcond");
1629 ExitingBlock->getInstList().insert(TermBr, Cond);
1631 // Clone the IVUse, as the old use still exists!
1632 CondUse = &IU.AddUser(CondUse->getExpr(),
1633 Cond, CondUse->getOperandValToReplace());
1634 TermBr->replaceUsesOfWith(OldCond, Cond);
1638 // If we get to here, we know that we can transform the setcc instruction to
1639 // use the post-incremented version of the IV, allowing us to coalesce the
1640 // live ranges for the IV correctly.
1641 CondUse->transformToPostInc(L);
1644 PostIncs.insert(Cond);
1648 // Determine an insertion point for the loop induction variable increment. It
1649 // must dominate all the post-inc comparisons we just set up, and it must
1650 // dominate the loop latch edge.
1651 IVIncInsertPos = L->getLoopLatch()->getTerminator();
1652 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(),
1653 E = PostIncs.end(); I != E; ++I) {
1655 DT.findNearestCommonDominator(IVIncInsertPos->getParent(),
1657 if (BB == (*I)->getParent())
1658 IVIncInsertPos = *I;
1659 else if (BB != IVIncInsertPos->getParent())
1660 IVIncInsertPos = BB->getTerminator();
1667 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset,
1668 LSRUse::KindType Kind, const Type *AccessTy) {
1669 int64_t NewMinOffset = LU.MinOffset;
1670 int64_t NewMaxOffset = LU.MaxOffset;
1671 const Type *NewAccessTy = AccessTy;
1673 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
1674 // something conservative, however this can pessimize in the case that one of
1675 // the uses will have all its uses outside the loop, for example.
1676 if (LU.Kind != Kind)
1678 // Conservatively assume HasBaseReg is true for now.
1679 if (NewOffset < LU.MinOffset) {
1680 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, /*HasBaseReg=*/true,
1681 Kind, AccessTy, TLI))
1683 NewMinOffset = NewOffset;
1684 } else if (NewOffset > LU.MaxOffset) {
1685 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, /*HasBaseReg=*/true,
1686 Kind, AccessTy, TLI))
1688 NewMaxOffset = NewOffset;
1690 // Check for a mismatched access type, and fall back conservatively as needed.
1691 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1692 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
1695 LU.MinOffset = NewMinOffset;
1696 LU.MaxOffset = NewMaxOffset;
1697 LU.AccessTy = NewAccessTy;
1698 if (NewOffset != LU.Offsets.back())
1699 LU.Offsets.push_back(NewOffset);
1703 /// getUse - Return an LSRUse index and an offset value for a fixup which
1704 /// needs the given expression, with the given kind and optional access type.
1705 /// Either reuse an existing use or create a new one, as needed.
1706 std::pair<size_t, int64_t>
1707 LSRInstance::getUse(const SCEV *&Expr,
1708 LSRUse::KindType Kind, const Type *AccessTy) {
1709 const SCEV *Copy = Expr;
1710 int64_t Offset = ExtractImmediate(Expr, SE);
1712 // Basic uses can't accept any offset, for example.
1713 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
1718 std::pair<UseMapTy::iterator, bool> P =
1719 UseMap.insert(std::make_pair(Expr, 0));
1721 // A use already existed with this base.
1722 size_t LUIdx = P.first->second;
1723 LSRUse &LU = Uses[LUIdx];
1724 if (reconcileNewOffset(LU, Offset, Kind, AccessTy))
1726 return std::make_pair(LUIdx, Offset);
1729 // Create a new use.
1730 size_t LUIdx = Uses.size();
1731 P.first->second = LUIdx;
1732 Uses.push_back(LSRUse(Kind, AccessTy));
1733 LSRUse &LU = Uses[LUIdx];
1735 // We don't need to track redundant offsets, but we don't need to go out
1736 // of our way here to avoid them.
1737 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1738 LU.Offsets.push_back(Offset);
1740 LU.MinOffset = Offset;
1741 LU.MaxOffset = Offset;
1742 return std::make_pair(LUIdx, Offset);
1745 void LSRInstance::CollectInterestingTypesAndFactors() {
1746 SmallSetVector<const SCEV *, 4> Strides;
1748 // Collect interesting types and strides.
1749 SmallVector<const SCEV *, 4> Worklist;
1750 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1751 const SCEV *Expr = UI->getExpr();
1753 // Collect interesting types.
1754 Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
1756 // Add strides for mentioned loops.
1757 Worklist.push_back(Expr);
1759 const SCEV *S = Worklist.pop_back_val();
1760 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
1761 Strides.insert(AR->getStepRecurrence(SE));
1762 Worklist.push_back(AR->getStart());
1763 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1764 Worklist.insert(Worklist.end(), Add->op_begin(), Add->op_end());
1766 } while (!Worklist.empty());
1769 // Compute interesting factors from the set of interesting strides.
1770 for (SmallSetVector<const SCEV *, 4>::const_iterator
1771 I = Strides.begin(), E = Strides.end(); I != E; ++I)
1772 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter =
1773 next(I); NewStrideIter != E; ++NewStrideIter) {
1774 const SCEV *OldStride = *I;
1775 const SCEV *NewStride = *NewStrideIter;
1777 if (SE.getTypeSizeInBits(OldStride->getType()) !=
1778 SE.getTypeSizeInBits(NewStride->getType())) {
1779 if (SE.getTypeSizeInBits(OldStride->getType()) >
1780 SE.getTypeSizeInBits(NewStride->getType()))
1781 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType());
1783 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType());
1785 if (const SCEVConstant *Factor =
1786 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride,
1788 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1789 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1790 } else if (const SCEVConstant *Factor =
1791 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride,
1794 if (Factor->getValue()->getValue().getMinSignedBits() <= 64)
1795 Factors.insert(Factor->getValue()->getValue().getSExtValue());
1799 // If all uses use the same type, don't bother looking for truncation-based
1801 if (Types.size() == 1)
1804 DEBUG(print_factors_and_types(dbgs()));
1807 void LSRInstance::CollectFixupsAndInitialFormulae() {
1808 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
1810 LSRFixup &LF = getNewFixup();
1811 LF.UserInst = UI->getUser();
1812 LF.OperandValToReplace = UI->getOperandValToReplace();
1813 LF.PostIncLoops = UI->getPostIncLoops();
1815 LSRUse::KindType Kind = LSRUse::Basic;
1816 const Type *AccessTy = 0;
1817 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) {
1818 Kind = LSRUse::Address;
1819 AccessTy = getAccessType(LF.UserInst);
1822 const SCEV *S = UI->getExpr();
1824 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
1825 // (N - i == 0), and this allows (N - i) to be the expression that we work
1826 // with rather than just N or i, so we can consider the register
1827 // requirements for both N and i at the same time. Limiting this code to
1828 // equality icmps is not a problem because all interesting loops use
1829 // equality icmps, thanks to IndVarSimplify.
1830 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst))
1831 if (CI->isEquality()) {
1832 // Swap the operands if needed to put the OperandValToReplace on the
1833 // left, for consistency.
1834 Value *NV = CI->getOperand(1);
1835 if (NV == LF.OperandValToReplace) {
1836 CI->setOperand(1, CI->getOperand(0));
1837 CI->setOperand(0, NV);
1840 // x == y --> x - y == 0
1841 const SCEV *N = SE.getSCEV(NV);
1842 if (N->isLoopInvariant(L)) {
1843 Kind = LSRUse::ICmpZero;
1844 S = SE.getMinusSCEV(N, S);
1847 // -1 and the negations of all interesting strides (except the negation
1848 // of -1) are now also interesting.
1849 for (size_t i = 0, e = Factors.size(); i != e; ++i)
1850 if (Factors[i] != -1)
1851 Factors.insert(-(uint64_t)Factors[i]);
1855 // Set up the initial formula for this use.
1856 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy);
1858 LF.Offset = P.second;
1859 LSRUse &LU = Uses[LF.LUIdx];
1860 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
1862 // If this is the first use of this LSRUse, give it a formula.
1863 if (LU.Formulae.empty()) {
1864 InsertInitialFormula(S, LU, LF.LUIdx);
1865 CountRegisters(LU.Formulae.back(), LF.LUIdx);
1869 DEBUG(print_fixups(dbgs()));
1873 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
1875 F.InitialMatch(S, L, SE, DT);
1876 bool Inserted = InsertFormula(LU, LUIdx, F);
1877 assert(Inserted && "Initial formula already exists!"); (void)Inserted;
1881 LSRInstance::InsertSupplementalFormula(const SCEV *S,
1882 LSRUse &LU, size_t LUIdx) {
1884 F.BaseRegs.push_back(S);
1885 F.AM.HasBaseReg = true;
1886 bool Inserted = InsertFormula(LU, LUIdx, F);
1887 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted;
1890 /// CountRegisters - Note which registers are used by the given formula,
1891 /// updating RegUses.
1892 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) {
1894 RegUses.CountRegister(F.ScaledReg, LUIdx);
1895 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
1896 E = F.BaseRegs.end(); I != E; ++I)
1897 RegUses.CountRegister(*I, LUIdx);
1900 /// InsertFormula - If the given formula has not yet been inserted, add it to
1901 /// the list, and return true. Return false otherwise.
1902 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) {
1903 if (!LU.InsertFormula(F))
1906 CountRegisters(F, LUIdx);
1910 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of
1911 /// loop-invariant values which we're tracking. These other uses will pin these
1912 /// values in registers, making them less profitable for elimination.
1913 /// TODO: This currently misses non-constant addrec step registers.
1914 /// TODO: Should this give more weight to users inside the loop?
1916 LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
1917 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end());
1918 SmallPtrSet<const SCEV *, 8> Inserted;
1920 while (!Worklist.empty()) {
1921 const SCEV *S = Worklist.pop_back_val();
1923 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S))
1924 Worklist.insert(Worklist.end(), N->op_begin(), N->op_end());
1925 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
1926 Worklist.push_back(C->getOperand());
1927 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) {
1928 Worklist.push_back(D->getLHS());
1929 Worklist.push_back(D->getRHS());
1930 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
1931 if (!Inserted.insert(U)) continue;
1932 const Value *V = U->getValue();
1933 if (const Instruction *Inst = dyn_cast<Instruction>(V))
1934 if (L->contains(Inst)) continue;
1935 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
1937 const Instruction *UserInst = dyn_cast<Instruction>(*UI);
1938 // Ignore non-instructions.
1941 // Ignore instructions in other functions (as can happen with
1943 if (UserInst->getParent()->getParent() != L->getHeader()->getParent())
1945 // Ignore instructions not dominated by the loop.
1946 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ?
1947 UserInst->getParent() :
1948 cast<PHINode>(UserInst)->getIncomingBlock(
1949 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()));
1950 if (!DT.dominates(L->getHeader(), UseBB))
1952 // Ignore uses which are part of other SCEV expressions, to avoid
1953 // analyzing them multiple times.
1954 if (SE.isSCEVable(UserInst->getType())) {
1955 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst));
1956 // If the user is a no-op, look through to its uses.
1957 if (!isa<SCEVUnknown>(UserS))
1961 SE.getUnknown(const_cast<Instruction *>(UserInst)));
1965 // Ignore icmp instructions which are already being analyzed.
1966 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) {
1967 unsigned OtherIdx = !UI.getOperandNo();
1968 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx));
1969 if (SE.getSCEV(OtherOp)->hasComputableLoopEvolution(L))
1973 LSRFixup &LF = getNewFixup();
1974 LF.UserInst = const_cast<Instruction *>(UserInst);
1975 LF.OperandValToReplace = UI.getUse();
1976 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0);
1978 LF.Offset = P.second;
1979 LSRUse &LU = Uses[LF.LUIdx];
1980 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
1981 InsertSupplementalFormula(U, LU, LF.LUIdx);
1982 CountRegisters(LU.Formulae.back(), Uses.size() - 1);
1989 /// CollectSubexprs - Split S into subexpressions which can be pulled out into
1990 /// separate registers. If C is non-null, multiply each subexpression by C.
1991 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C,
1992 SmallVectorImpl<const SCEV *> &Ops,
1993 ScalarEvolution &SE) {
1994 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
1995 // Break out add operands.
1996 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
1998 CollectSubexprs(*I, C, Ops, SE);
2000 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
2001 // Split a non-zero base out of an addrec.
2002 if (!AR->getStart()->isZero()) {
2003 CollectSubexprs(SE.getAddRecExpr(SE.getIntegerSCEV(0, AR->getType()),
2004 AR->getStepRecurrence(SE),
2005 AR->getLoop()), C, Ops, SE);
2006 CollectSubexprs(AR->getStart(), C, Ops, SE);
2009 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) {
2010 // Break (C * (a + b + c)) into C*a + C*b + C*c.
2011 if (Mul->getNumOperands() == 2)
2012 if (const SCEVConstant *Op0 =
2013 dyn_cast<SCEVConstant>(Mul->getOperand(0))) {
2014 CollectSubexprs(Mul->getOperand(1),
2015 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0,
2021 // Otherwise use the value itself.
2022 Ops.push_back(C ? SE.getMulExpr(C, S) : S);
2025 /// GenerateReassociations - Split out subexpressions from adds and the bases of
2027 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx,
2030 // Arbitrarily cap recursion to protect compile time.
2031 if (Depth >= 3) return;
2033 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2034 const SCEV *BaseReg = Base.BaseRegs[i];
2036 SmallVector<const SCEV *, 8> AddOps;
2037 CollectSubexprs(BaseReg, 0, AddOps, SE);
2038 if (AddOps.size() == 1) continue;
2040 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(),
2041 JE = AddOps.end(); J != JE; ++J) {
2042 // Don't pull a constant into a register if the constant could be folded
2043 // into an immediate field.
2044 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
2045 Base.getNumRegs() > 1,
2046 LU.Kind, LU.AccessTy, TLI, SE))
2049 // Collect all operands except *J.
2050 SmallVector<const SCEV *, 8> InnerAddOps;
2051 for (SmallVectorImpl<const SCEV *>::const_iterator K = AddOps.begin(),
2052 KE = AddOps.end(); K != KE; ++K)
2054 InnerAddOps.push_back(*K);
2056 // Don't leave just a constant behind in a register if the constant could
2057 // be folded into an immediate field.
2058 if (InnerAddOps.size() == 1 &&
2059 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
2060 Base.getNumRegs() > 1,
2061 LU.Kind, LU.AccessTy, TLI, SE))
2065 F.BaseRegs[i] = SE.getAddExpr(InnerAddOps);
2066 F.BaseRegs.push_back(*J);
2067 if (InsertFormula(LU, LUIdx, F))
2068 // If that formula hadn't been seen before, recurse to find more like
2070 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1);
2075 /// GenerateCombinations - Generate a formula consisting of all of the
2076 /// loop-dominating registers added into a single register.
2077 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx,
2079 // This method is only interesting on a plurality of registers.
2080 if (Base.BaseRegs.size() <= 1) return;
2084 SmallVector<const SCEV *, 4> Ops;
2085 for (SmallVectorImpl<const SCEV *>::const_iterator
2086 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) {
2087 const SCEV *BaseReg = *I;
2088 if (BaseReg->properlyDominates(L->getHeader(), &DT) &&
2089 !BaseReg->hasComputableLoopEvolution(L))
2090 Ops.push_back(BaseReg);
2092 F.BaseRegs.push_back(BaseReg);
2094 if (Ops.size() > 1) {
2095 const SCEV *Sum = SE.getAddExpr(Ops);
2096 // TODO: If Sum is zero, it probably means ScalarEvolution missed an
2097 // opportunity to fold something. For now, just ignore such cases
2098 // rather than proceed with zero in a register.
2099 if (!Sum->isZero()) {
2100 F.BaseRegs.push_back(Sum);
2101 (void)InsertFormula(LU, LUIdx, F);
2106 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets.
2107 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx,
2109 // We can't add a symbolic offset if the address already contains one.
2110 if (Base.AM.BaseGV) return;
2112 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2113 const SCEV *G = Base.BaseRegs[i];
2114 GlobalValue *GV = ExtractSymbol(G, SE);
2115 if (G->isZero() || !GV)
2119 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2120 LU.Kind, LU.AccessTy, TLI))
2123 (void)InsertFormula(LU, LUIdx, F);
2127 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets.
2128 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx,
2130 // TODO: For now, just add the min and max offset, because it usually isn't
2131 // worthwhile looking at everything inbetween.
2132 SmallVector<int64_t, 4> Worklist;
2133 Worklist.push_back(LU.MinOffset);
2134 if (LU.MaxOffset != LU.MinOffset)
2135 Worklist.push_back(LU.MaxOffset);
2137 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
2138 const SCEV *G = Base.BaseRegs[i];
2140 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(),
2141 E = Worklist.end(); I != E; ++I) {
2143 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
2144 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
2145 LU.Kind, LU.AccessTy, TLI)) {
2146 F.BaseRegs[i] = SE.getAddExpr(G, SE.getIntegerSCEV(*I, G->getType()));
2148 (void)InsertFormula(LU, LUIdx, F);
2152 int64_t Imm = ExtractImmediate(G, SE);
2153 if (G->isZero() || Imm == 0)
2156 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
2157 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
2158 LU.Kind, LU.AccessTy, TLI))
2161 (void)InsertFormula(LU, LUIdx, F);
2165 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up
2166 /// the comparison. For example, x == y -> x*c == y*c.
2167 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx,
2169 if (LU.Kind != LSRUse::ICmpZero) return;
2171 // Determine the integer type for the base formula.
2172 const Type *IntTy = Base.getType();
2174 if (SE.getTypeSizeInBits(IntTy) > 64) return;
2176 // Don't do this if there is more than one offset.
2177 if (LU.MinOffset != LU.MaxOffset) return;
2179 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!");
2181 // Check each interesting stride.
2182 for (SmallSetVector<int64_t, 8>::const_iterator
2183 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2184 int64_t Factor = *I;
2187 // Check that the multiplication doesn't overflow.
2188 if (F.AM.BaseOffs == INT64_MIN && Factor == -1)
2190 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs * Factor;
2191 if (F.AM.BaseOffs / Factor != Base.AM.BaseOffs)
2194 // Check that multiplying with the use offset doesn't overflow.
2195 int64_t Offset = LU.MinOffset;
2196 if (Offset == INT64_MIN && Factor == -1)
2198 Offset = (uint64_t)Offset * Factor;
2199 if (Offset / Factor != LU.MinOffset)
2202 // Check that this scale is legal.
2203 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
2206 // Compensate for the use having MinOffset built into it.
2207 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset;
2209 const SCEV *FactorS = SE.getIntegerSCEV(Factor, IntTy);
2211 // Check that multiplying with each base register doesn't overflow.
2212 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) {
2213 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS);
2214 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i])
2218 // Check that multiplying with the scaled register doesn't overflow.
2220 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS);
2221 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg)
2225 // If we make it here and it's legal, add it.
2226 (void)InsertFormula(LU, LUIdx, F);
2231 /// GenerateScales - Generate stride factor reuse formulae by making use of
2232 /// scaled-offset address modes, for example.
2233 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx,
2235 // Determine the integer type for the base formula.
2236 const Type *IntTy = Base.getType();
2239 // If this Formula already has a scaled register, we can't add another one.
2240 if (Base.AM.Scale != 0) return;
2242 // Check each interesting stride.
2243 for (SmallSetVector<int64_t, 8>::const_iterator
2244 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
2245 int64_t Factor = *I;
2247 Base.AM.Scale = Factor;
2248 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
2249 // Check whether this scale is going to be legal.
2250 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2251 LU.Kind, LU.AccessTy, TLI)) {
2252 // As a special-case, handle special out-of-loop Basic users specially.
2253 // TODO: Reconsider this special case.
2254 if (LU.Kind == LSRUse::Basic &&
2255 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
2256 LSRUse::Special, LU.AccessTy, TLI) &&
2257 LU.AllFixupsOutsideLoop)
2258 LU.Kind = LSRUse::Special;
2262 // For an ICmpZero, negating a solitary base register won't lead to
2264 if (LU.Kind == LSRUse::ICmpZero &&
2265 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV)
2267 // For each addrec base reg, apply the scale, if possible.
2268 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i)
2269 if (const SCEVAddRecExpr *AR =
2270 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) {
2271 const SCEV *FactorS = SE.getIntegerSCEV(Factor, IntTy);
2272 if (FactorS->isZero())
2274 // Divide out the factor, ignoring high bits, since we'll be
2275 // scaling the value back up in the end.
2276 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
2277 // TODO: This could be optimized to avoid all the copying.
2279 F.ScaledReg = Quotient;
2280 std::swap(F.BaseRegs[i], F.BaseRegs.back());
2281 F.BaseRegs.pop_back();
2282 (void)InsertFormula(LU, LUIdx, F);
2288 /// GenerateTruncates - Generate reuse formulae from different IV types.
2289 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx,
2291 // This requires TargetLowering to tell us which truncates are free.
2294 // Don't bother truncating symbolic values.
2295 if (Base.AM.BaseGV) return;
2297 // Determine the integer type for the base formula.
2298 const Type *DstTy = Base.getType();
2300 DstTy = SE.getEffectiveSCEVType(DstTy);
2302 for (SmallSetVector<const Type *, 4>::const_iterator
2303 I = Types.begin(), E = Types.end(); I != E; ++I) {
2304 const Type *SrcTy = *I;
2305 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
2308 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
2309 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(),
2310 JE = F.BaseRegs.end(); J != JE; ++J)
2311 *J = SE.getAnyExtendExpr(*J, SrcTy);
2313 // TODO: This assumes we've done basic processing on all uses and
2314 // have an idea what the register usage is.
2315 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses))
2318 (void)InsertFormula(LU, LUIdx, F);
2325 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to
2326 /// defer modifications so that the search phase doesn't have to worry about
2327 /// the data structures moving underneath it.
2331 const SCEV *OrigReg;
2333 WorkItem(size_t LI, int64_t I, const SCEV *R)
2334 : LUIdx(LI), Imm(I), OrigReg(R) {}
2336 void print(raw_ostream &OS) const;
2342 void WorkItem::print(raw_ostream &OS) const {
2343 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx
2344 << " , add offset " << Imm;
2347 void WorkItem::dump() const {
2348 print(errs()); errs() << '\n';
2351 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant
2352 /// distance apart and try to form reuse opportunities between them.
2353 void LSRInstance::GenerateCrossUseConstantOffsets() {
2354 // Group the registers by their value without any added constant offset.
2355 typedef std::map<int64_t, const SCEV *> ImmMapTy;
2356 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy;
2358 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap;
2359 SmallVector<const SCEV *, 8> Sequence;
2360 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2362 const SCEV *Reg = *I;
2363 int64_t Imm = ExtractImmediate(Reg, SE);
2364 std::pair<RegMapTy::iterator, bool> Pair =
2365 Map.insert(std::make_pair(Reg, ImmMapTy()));
2367 Sequence.push_back(Reg);
2368 Pair.first->second.insert(std::make_pair(Imm, *I));
2369 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I);
2372 // Now examine each set of registers with the same base value. Build up
2373 // a list of work to do and do the work in a separate step so that we're
2374 // not adding formulae and register counts while we're searching.
2375 SmallVector<WorkItem, 32> WorkItems;
2376 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
2377 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
2378 E = Sequence.end(); I != E; ++I) {
2379 const SCEV *Reg = *I;
2380 const ImmMapTy &Imms = Map.find(Reg)->second;
2382 // It's not worthwhile looking for reuse if there's only one offset.
2383 if (Imms.size() == 1)
2386 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':';
2387 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2389 dbgs() << ' ' << J->first;
2392 // Examine each offset.
2393 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end();
2395 const SCEV *OrigReg = J->second;
2397 int64_t JImm = J->first;
2398 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg);
2400 if (!isa<SCEVConstant>(OrigReg) &&
2401 UsedByIndicesMap[Reg].count() == 1) {
2402 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n');
2406 // Conservatively examine offsets between this orig reg a few selected
2408 ImmMapTy::const_iterator OtherImms[] = {
2409 Imms.begin(), prior(Imms.end()),
2410 Imms.upper_bound((Imms.begin()->first + prior(Imms.end())->first) / 2)
2412 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) {
2413 ImmMapTy::const_iterator M = OtherImms[i];
2414 if (M == J || M == JE) continue;
2416 // Compute the difference between the two.
2417 int64_t Imm = (uint64_t)JImm - M->first;
2418 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2419 LUIdx = UsedByIndices.find_next(LUIdx))
2420 // Make a memo of this use, offset, and register tuple.
2421 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2422 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
2429 UsedByIndicesMap.clear();
2430 UniqueItems.clear();
2432 // Now iterate through the worklist and add new formulae.
2433 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
2434 E = WorkItems.end(); I != E; ++I) {
2435 const WorkItem &WI = *I;
2436 size_t LUIdx = WI.LUIdx;
2437 LSRUse &LU = Uses[LUIdx];
2438 int64_t Imm = WI.Imm;
2439 const SCEV *OrigReg = WI.OrigReg;
2441 const Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType());
2442 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm));
2443 unsigned BitWidth = SE.getTypeSizeInBits(IntTy);
2445 // TODO: Use a more targeted data structure.
2446 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) {
2447 Formula F = LU.Formulae[L];
2448 // Use the immediate in the scaled register.
2449 if (F.ScaledReg == OrigReg) {
2450 int64_t Offs = (uint64_t)F.AM.BaseOffs +
2451 Imm * (uint64_t)F.AM.Scale;
2452 // Don't create 50 + reg(-50).
2453 if (F.referencesReg(SE.getSCEV(
2454 ConstantInt::get(IntTy, -(uint64_t)Offs))))
2457 NewF.AM.BaseOffs = Offs;
2458 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2459 LU.Kind, LU.AccessTy, TLI))
2461 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
2463 // If the new scale is a constant in a register, and adding the constant
2464 // value to the immediate would produce a value closer to zero than the
2465 // immediate itself, then the formula isn't worthwhile.
2466 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg))
2467 if (C->getValue()->getValue().isNegative() !=
2468 (NewF.AM.BaseOffs < 0) &&
2469 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale))
2470 .ule(abs64(NewF.AM.BaseOffs)))
2474 (void)InsertFormula(LU, LUIdx, NewF);
2476 // Use the immediate in a base register.
2477 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) {
2478 const SCEV *BaseReg = F.BaseRegs[N];
2479 if (BaseReg != OrigReg)
2482 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
2483 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
2484 LU.Kind, LU.AccessTy, TLI))
2486 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg);
2488 // If the new formula has a constant in a register, and adding the
2489 // constant value to the immediate would produce a value closer to
2490 // zero than the immediate itself, then the formula isn't worthwhile.
2491 for (SmallVectorImpl<const SCEV *>::const_iterator
2492 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end();
2494 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J))
2495 if (C->getValue()->getValue().isNegative() !=
2496 (NewF.AM.BaseOffs < 0) &&
2497 C->getValue()->getValue().abs()
2498 .ule(abs64(NewF.AM.BaseOffs)))
2502 (void)InsertFormula(LU, LUIdx, NewF);
2511 /// GenerateAllReuseFormulae - Generate formulae for each use.
2513 LSRInstance::GenerateAllReuseFormulae() {
2514 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan
2515 // queries are more precise.
2516 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2517 LSRUse &LU = Uses[LUIdx];
2518 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2519 GenerateReassociations(LU, LUIdx, LU.Formulae[i]);
2520 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2521 GenerateCombinations(LU, LUIdx, LU.Formulae[i]);
2523 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2524 LSRUse &LU = Uses[LUIdx];
2525 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2526 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]);
2527 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2528 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]);
2529 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2530 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]);
2531 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2532 GenerateScales(LU, LUIdx, LU.Formulae[i]);
2534 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2535 LSRUse &LU = Uses[LUIdx];
2536 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i)
2537 GenerateTruncates(LU, LUIdx, LU.Formulae[i]);
2540 GenerateCrossUseConstantOffsets();
2543 /// If their are multiple formulae with the same set of registers used
2544 /// by other uses, pick the best one and delete the others.
2545 void LSRInstance::FilterOutUndesirableDedicatedRegisters() {
2547 bool Changed = false;
2550 // Collect the best formula for each unique set of shared registers. This
2551 // is reset for each use.
2552 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo>
2554 BestFormulaeTy BestFormulae;
2556 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
2557 LSRUse &LU = Uses[LUIdx];
2558 FormulaSorter Sorter(L, LU, SE, DT);
2560 // Clear out the set of used regs; it will be recomputed.
2563 for (size_t FIdx = 0, NumForms = LU.Formulae.size();
2564 FIdx != NumForms; ++FIdx) {
2565 Formula &F = LU.Formulae[FIdx];
2567 SmallVector<const SCEV *, 2> Key;
2568 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(),
2569 JE = F.BaseRegs.end(); J != JE; ++J) {
2570 const SCEV *Reg = *J;
2571 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx))
2575 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx))
2576 Key.push_back(F.ScaledReg);
2577 // Unstable sort by host order ok, because this is only used for
2579 std::sort(Key.begin(), Key.end());
2581 std::pair<BestFormulaeTy::const_iterator, bool> P =
2582 BestFormulae.insert(std::make_pair(Key, FIdx));
2584 Formula &Best = LU.Formulae[P.first->second];
2585 if (Sorter.operator()(F, Best))
2587 DEBUG(dbgs() << "Filtering out "; F.print(dbgs());
2589 " in favor of "; Best.print(dbgs());
2594 std::swap(F, LU.Formulae.back());
2595 LU.Formulae.pop_back();
2600 if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
2601 LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
2603 BestFormulae.clear();
2606 DEBUG(if (Changed) {
2608 "After filtering out undesirable candidates:\n";
2613 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of
2614 /// formulae to choose from, use some rough heuristics to prune down the number
2615 /// of formulae. This keeps the main solver from taking an extraordinary amount
2616 /// of time in some worst-case scenarios.
2617 void LSRInstance::NarrowSearchSpaceUsingHeuristics() {
2618 // This is a rough guess that seems to work fairly well.
2619 const size_t Limit = UINT16_MAX;
2621 SmallPtrSet<const SCEV *, 4> Taken;
2623 // Estimate the worst-case number of solutions we might consider. We almost
2624 // never consider this many solutions because we prune the search space,
2625 // but the pruning isn't always sufficient.
2627 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
2628 E = Uses.end(); I != E; ++I) {
2629 size_t FSize = I->Formulae.size();
2630 if (FSize >= Limit) {
2641 // Ok, we have too many of formulae on our hands to conveniently handle.
2642 // Use a rough heuristic to thin out the list.
2644 // Pick the register which is used by the most LSRUses, which is likely
2645 // to be a good reuse register candidate.
2646 const SCEV *Best = 0;
2647 unsigned BestNum = 0;
2648 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end();
2650 const SCEV *Reg = *I;
2651 if (Taken.count(Reg))
2656 unsigned Count = RegUses.getUsedByIndices(Reg).count();
2657 if (Count > BestNum) {
2664 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best
2665 << " will yield profitable reuse.\n");
2668 // In any use with formulae which references this register, delete formulae
2669 // which don't reference it.
2670 for (SmallVectorImpl<LSRUse>::iterator I = Uses.begin(),
2671 E = Uses.end(); I != E; ++I) {
2673 if (!LU.Regs.count(Best)) continue;
2675 // Clear out the set of used regs; it will be recomputed.
2678 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) {
2679 Formula &F = LU.Formulae[i];
2680 if (!F.referencesReg(Best)) {
2681 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n');
2682 std::swap(LU.Formulae.back(), F);
2683 LU.Formulae.pop_back();
2689 if (F.ScaledReg) LU.Regs.insert(F.ScaledReg);
2690 LU.Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end());
2694 DEBUG(dbgs() << "After pre-selection:\n";
2695 print_uses(dbgs()));
2699 /// SolveRecurse - This is the recursive solver.
2700 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution,
2702 SmallVectorImpl<const Formula *> &Workspace,
2703 const Cost &CurCost,
2704 const SmallPtrSet<const SCEV *, 16> &CurRegs,
2705 DenseSet<const SCEV *> &VisitedRegs) const {
2708 // - use more aggressive filtering
2709 // - sort the formula so that the most profitable solutions are found first
2710 // - sort the uses too
2712 // - don't compute a cost, and then compare. compare while computing a cost
2714 // - track register sets with SmallBitVector
2716 const LSRUse &LU = Uses[Workspace.size()];
2718 // If this use references any register that's already a part of the
2719 // in-progress solution, consider it a requirement that a formula must
2720 // reference that register in order to be considered. This prunes out
2721 // unprofitable searching.
2722 SmallSetVector<const SCEV *, 4> ReqRegs;
2723 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(),
2724 E = CurRegs.end(); I != E; ++I)
2725 if (LU.Regs.count(*I))
2728 bool AnySatisfiedReqRegs = false;
2729 SmallPtrSet<const SCEV *, 16> NewRegs;
2732 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(),
2733 E = LU.Formulae.end(); I != E; ++I) {
2734 const Formula &F = *I;
2736 // Ignore formulae which do not use any of the required registers.
2737 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(),
2738 JE = ReqRegs.end(); J != JE; ++J) {
2739 const SCEV *Reg = *J;
2740 if ((!F.ScaledReg || F.ScaledReg != Reg) &&
2741 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) ==
2745 AnySatisfiedReqRegs = true;
2747 // Evaluate the cost of the current formula. If it's already worse than
2748 // the current best, prune the search at that point.
2751 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT);
2752 if (NewCost < SolutionCost) {
2753 Workspace.push_back(&F);
2754 if (Workspace.size() != Uses.size()) {
2755 SolveRecurse(Solution, SolutionCost, Workspace, NewCost,
2756 NewRegs, VisitedRegs);
2757 if (F.getNumRegs() == 1 && Workspace.size() == 1)
2758 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]);
2760 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs());
2761 dbgs() << ". Regs:";
2762 for (SmallPtrSet<const SCEV *, 16>::const_iterator
2763 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I)
2764 dbgs() << ' ' << **I;
2767 SolutionCost = NewCost;
2768 Solution = Workspace;
2770 Workspace.pop_back();
2775 // If none of the formulae had all of the required registers, relax the
2776 // constraint so that we don't exclude all formulae.
2777 if (!AnySatisfiedReqRegs) {
2783 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const {
2784 SmallVector<const Formula *, 8> Workspace;
2786 SolutionCost.Loose();
2788 SmallPtrSet<const SCEV *, 16> CurRegs;
2789 DenseSet<const SCEV *> VisitedRegs;
2790 Workspace.reserve(Uses.size());
2792 SolveRecurse(Solution, SolutionCost, Workspace, CurCost,
2793 CurRegs, VisitedRegs);
2795 // Ok, we've now made all our decisions.
2796 DEBUG(dbgs() << "\n"
2797 "The chosen solution requires "; SolutionCost.print(dbgs());
2799 for (size_t i = 0, e = Uses.size(); i != e; ++i) {
2801 Uses[i].print(dbgs());
2804 Solution[i]->print(dbgs());
2809 /// getImmediateDominator - A handy utility for the specific DominatorTree
2810 /// query that we need here.
2812 static BasicBlock *getImmediateDominator(BasicBlock *BB, DominatorTree &DT) {
2813 DomTreeNode *Node = DT.getNode(BB);
2814 if (!Node) return 0;
2815 Node = Node->getIDom();
2816 if (!Node) return 0;
2817 return Node->getBlock();
2820 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up
2821 /// the dominator tree far as we can go while still being dominated by the
2822 /// input positions. This helps canonicalize the insert position, which
2823 /// encourages sharing.
2824 BasicBlock::iterator
2825 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
2826 const SmallVectorImpl<Instruction *> &Inputs)
2829 const Loop *IPLoop = LI.getLoopFor(IP->getParent());
2830 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
2833 for (BasicBlock *Rung = IP->getParent(); ; Rung = IDom) {
2834 IDom = getImmediateDominator(Rung, DT);
2835 if (!IDom) return IP;
2837 // Don't climb into a loop though.
2838 const Loop *IDomLoop = LI.getLoopFor(IDom);
2839 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
2840 if (IDomDepth <= IPLoopDepth &&
2841 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
2845 bool AllDominate = true;
2846 Instruction *BetterPos = 0;
2847 Instruction *Tentative = IDom->getTerminator();
2848 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(),
2849 E = Inputs.end(); I != E; ++I) {
2850 Instruction *Inst = *I;
2851 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) {
2852 AllDominate = false;
2855 // Attempt to find an insert position in the middle of the block,
2856 // instead of at the end, so that it can be used for other expansions.
2857 if (IDom == Inst->getParent() &&
2858 (!BetterPos || DT.dominates(BetterPos, Inst)))
2859 BetterPos = next(BasicBlock::iterator(Inst));
2872 /// AdjustInsertPositionForExpand - Determine an input position which will be
2873 /// dominated by the operands and which will dominate the result.
2874 BasicBlock::iterator
2875 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP,
2877 const LSRUse &LU) const {
2878 // Collect some instructions which must be dominated by the
2879 // expanding replacement. These must be dominated by any operands that
2880 // will be required in the expansion.
2881 SmallVector<Instruction *, 4> Inputs;
2882 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
2883 Inputs.push_back(I);
2884 if (LU.Kind == LSRUse::ICmpZero)
2885 if (Instruction *I =
2886 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
2887 Inputs.push_back(I);
2888 if (LF.PostIncLoops.count(L)) {
2889 if (LF.isUseFullyOutsideLoop(L))
2890 Inputs.push_back(L->getLoopLatch()->getTerminator());
2892 Inputs.push_back(IVIncInsertPos);
2894 // The expansion must also be dominated by the increment positions of any
2895 // loops it for which it is using post-inc mode.
2896 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(),
2897 E = LF.PostIncLoops.end(); I != E; ++I) {
2898 const Loop *PIL = *I;
2899 if (PIL == L) continue;
2901 // Be dominated by the loop exit.
2902 SmallVector<BasicBlock *, 4> ExitingBlocks;
2903 PIL->getExitingBlocks(ExitingBlocks);
2904 if (!ExitingBlocks.empty()) {
2905 BasicBlock *BB = ExitingBlocks[0];
2906 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i)
2907 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]);
2908 Inputs.push_back(BB->getTerminator());
2912 // Then, climb up the immediate dominator tree as far as we can go while
2913 // still being dominated by the input positions.
2914 IP = HoistInsertPosition(IP, Inputs);
2916 // Don't insert instructions before PHI nodes.
2917 while (isa<PHINode>(IP)) ++IP;
2919 // Ignore debug intrinsics.
2920 while (isa<DbgInfoIntrinsic>(IP)) ++IP;
2925 Value *LSRInstance::Expand(const LSRFixup &LF,
2927 BasicBlock::iterator IP,
2928 SCEVExpander &Rewriter,
2929 SmallVectorImpl<WeakVH> &DeadInsts) const {
2930 const LSRUse &LU = Uses[LF.LUIdx];
2932 // Determine an input position which will be dominated by the operands and
2933 // which will dominate the result.
2934 IP = AdjustInsertPositionForExpand(IP, LF, LU);
2936 // Inform the Rewriter if we have a post-increment use, so that it can
2937 // perform an advantageous expansion.
2938 Rewriter.setPostInc(LF.PostIncLoops);
2940 // This is the type that the user actually needs.
2941 const Type *OpTy = LF.OperandValToReplace->getType();
2942 // This will be the type that we'll initially expand to.
2943 const Type *Ty = F.getType();
2945 // No type known; just expand directly to the ultimate type.
2947 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy))
2948 // Expand directly to the ultimate type if it's the right size.
2950 // This is the type to do integer arithmetic in.
2951 const Type *IntTy = SE.getEffectiveSCEVType(Ty);
2953 // Build up a list of operands to add together to form the full base.
2954 SmallVector<const SCEV *, 8> Ops;
2956 // Expand the BaseRegs portion.
2957 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(),
2958 E = F.BaseRegs.end(); I != E; ++I) {
2959 const SCEV *Reg = *I;
2960 assert(!Reg->isZero() && "Zero allocated in a base register!");
2962 // If we're expanding for a post-inc user, make the post-inc adjustment.
2963 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
2964 Reg = TransformForPostIncUse(Denormalize, Reg,
2965 LF.UserInst, LF.OperandValToReplace,
2968 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
2971 // Flush the operand list to suppress SCEVExpander hoisting.
2973 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
2975 Ops.push_back(SE.getUnknown(FullV));
2978 // Expand the ScaledReg portion.
2979 Value *ICmpScaledV = 0;
2980 if (F.AM.Scale != 0) {
2981 const SCEV *ScaledS = F.ScaledReg;
2983 // If we're expanding for a post-inc user, make the post-inc adjustment.
2984 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
2985 ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
2986 LF.UserInst, LF.OperandValToReplace,
2989 if (LU.Kind == LSRUse::ICmpZero) {
2990 // An interesting way of "folding" with an icmp is to use a negated
2991 // scale, which we'll implement by inserting it into the other operand
2993 assert(F.AM.Scale == -1 &&
2994 "The only scale supported by ICmpZero uses is -1!");
2995 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP);
2997 // Otherwise just expand the scaled register and an explicit scale,
2998 // which is expected to be matched as part of the address.
2999 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP));
3000 ScaledS = SE.getMulExpr(ScaledS,
3001 SE.getIntegerSCEV(F.AM.Scale,
3002 ScaledS->getType()));
3003 Ops.push_back(ScaledS);
3005 // Flush the operand list to suppress SCEVExpander hoisting.
3006 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3008 Ops.push_back(SE.getUnknown(FullV));
3012 // Expand the GV portion.
3014 Ops.push_back(SE.getUnknown(F.AM.BaseGV));
3016 // Flush the operand list to suppress SCEVExpander hoisting.
3017 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP);
3019 Ops.push_back(SE.getUnknown(FullV));
3022 // Expand the immediate portion.
3023 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset;
3025 if (LU.Kind == LSRUse::ICmpZero) {
3026 // The other interesting way of "folding" with an ICmpZero is to use a
3027 // negated immediate.
3029 ICmpScaledV = ConstantInt::get(IntTy, -Offset);
3031 Ops.push_back(SE.getUnknown(ICmpScaledV));
3032 ICmpScaledV = ConstantInt::get(IntTy, Offset);
3035 // Just add the immediate values. These again are expected to be matched
3036 // as part of the address.
3037 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset)));
3041 // Emit instructions summing all the operands.
3042 const SCEV *FullS = Ops.empty() ?
3043 SE.getIntegerSCEV(0, IntTy) :
3045 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
3047 // We're done expanding now, so reset the rewriter.
3048 Rewriter.clearPostInc();
3050 // An ICmpZero Formula represents an ICmp which we're handling as a
3051 // comparison against zero. Now that we've expanded an expression for that
3052 // form, update the ICmp's other operand.
3053 if (LU.Kind == LSRUse::ICmpZero) {
3054 ICmpInst *CI = cast<ICmpInst>(LF.UserInst);
3055 DeadInsts.push_back(CI->getOperand(1));
3056 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and "
3057 "a scale at the same time!");
3058 if (F.AM.Scale == -1) {
3059 if (ICmpScaledV->getType() != OpTy) {
3061 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false,
3063 ICmpScaledV, OpTy, "tmp", CI);
3066 CI->setOperand(1, ICmpScaledV);
3068 assert(F.AM.Scale == 0 &&
3069 "ICmp does not support folding a global value and "
3070 "a scale at the same time!");
3071 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy),
3073 if (C->getType() != OpTy)
3074 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false,
3078 CI->setOperand(1, C);
3085 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use
3086 /// of their operands effectively happens in their predecessor blocks, so the
3087 /// expression may need to be expanded in multiple places.
3088 void LSRInstance::RewriteForPHI(PHINode *PN,
3091 SCEVExpander &Rewriter,
3092 SmallVectorImpl<WeakVH> &DeadInsts,
3094 DenseMap<BasicBlock *, Value *> Inserted;
3095 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3096 if (PN->getIncomingValue(i) == LF.OperandValToReplace) {
3097 BasicBlock *BB = PN->getIncomingBlock(i);
3099 // If this is a critical edge, split the edge so that we do not insert
3100 // the code on all predecessor/successor paths. We do this unless this
3101 // is the canonical backedge for this loop, which complicates post-inc
3103 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 &&
3104 !isa<IndirectBrInst>(BB->getTerminator()) &&
3105 (PN->getParent() != L->getHeader() || !L->contains(BB))) {
3106 // Split the critical edge.
3107 BasicBlock *NewBB = SplitCriticalEdge(BB, PN->getParent(), P);
3109 // If PN is outside of the loop and BB is in the loop, we want to
3110 // move the block to be immediately before the PHI block, not
3111 // immediately after BB.
3112 if (L->contains(BB) && !L->contains(PN))
3113 NewBB->moveBefore(PN->getParent());
3115 // Splitting the edge can reduce the number of PHI entries we have.
3116 e = PN->getNumIncomingValues();
3118 i = PN->getBasicBlockIndex(BB);
3121 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair =
3122 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0)));
3124 PN->setIncomingValue(i, Pair.first->second);
3126 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts);
3128 // If this is reuse-by-noop-cast, insert the noop cast.
3129 const Type *OpTy = LF.OperandValToReplace->getType();
3130 if (FullV->getType() != OpTy)
3132 CastInst::Create(CastInst::getCastOpcode(FullV, false,
3134 FullV, LF.OperandValToReplace->getType(),
3135 "tmp", BB->getTerminator());
3137 PN->setIncomingValue(i, FullV);
3138 Pair.first->second = FullV;
3143 /// Rewrite - Emit instructions for the leading candidate expression for this
3144 /// LSRUse (this is called "expanding"), and update the UserInst to reference
3145 /// the newly expanded value.
3146 void LSRInstance::Rewrite(const LSRFixup &LF,
3148 SCEVExpander &Rewriter,
3149 SmallVectorImpl<WeakVH> &DeadInsts,
3151 // First, find an insertion point that dominates UserInst. For PHI nodes,
3152 // find the nearest block which dominates all the relevant uses.
3153 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) {
3154 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P);
3156 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts);
3158 // If this is reuse-by-noop-cast, insert the noop cast.
3159 const Type *OpTy = LF.OperandValToReplace->getType();
3160 if (FullV->getType() != OpTy) {
3162 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false),
3163 FullV, OpTy, "tmp", LF.UserInst);
3167 // Update the user. ICmpZero is handled specially here (for now) because
3168 // Expand may have updated one of the operands of the icmp already, and
3169 // its new value may happen to be equal to LF.OperandValToReplace, in
3170 // which case doing replaceUsesOfWith leads to replacing both operands
3171 // with the same value. TODO: Reorganize this.
3172 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero)
3173 LF.UserInst->setOperand(0, FullV);
3175 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV);
3178 DeadInsts.push_back(LF.OperandValToReplace);
3182 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution,
3184 // Keep track of instructions we may have made dead, so that
3185 // we can remove them after we are done working.
3186 SmallVector<WeakVH, 16> DeadInsts;
3188 SCEVExpander Rewriter(SE);
3189 Rewriter.disableCanonicalMode();
3190 Rewriter.setIVIncInsertPos(L, IVIncInsertPos);
3192 // Expand the new value definitions and update the users.
3193 for (size_t i = 0, e = Fixups.size(); i != e; ++i) {
3194 size_t LUIdx = Fixups[i].LUIdx;
3196 Rewrite(Fixups[i], *Solution[LUIdx], Rewriter, DeadInsts, P);
3201 // Clean up after ourselves. This must be done before deleting any
3205 Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
3208 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
3209 : IU(P->getAnalysis<IVUsers>()),
3210 SE(P->getAnalysis<ScalarEvolution>()),
3211 DT(P->getAnalysis<DominatorTree>()),
3212 LI(P->getAnalysis<LoopInfo>()),
3213 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
3215 // If LoopSimplify form is not available, stay out of trouble.
3216 if (!L->isLoopSimplifyForm()) return;
3218 // If there's no interesting work to be done, bail early.
3219 if (IU.empty()) return;
3221 DEBUG(dbgs() << "\nLSR on loop ";
3222 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
3225 /// OptimizeShadowIV - If IV is used in a int-to-float cast
3226 /// inside the loop then try to eliminate the cast operation.
3229 // Change loop terminating condition to use the postinc iv when possible.
3230 Changed |= OptimizeLoopTermCond();
3232 CollectInterestingTypesAndFactors();
3233 CollectFixupsAndInitialFormulae();
3234 CollectLoopInvariantFixupsAndFormulae();
3236 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n";
3237 print_uses(dbgs()));
3239 // Now use the reuse data to generate a bunch of interesting ways
3240 // to formulate the values needed for the uses.
3241 GenerateAllReuseFormulae();
3243 DEBUG(dbgs() << "\n"
3244 "After generating reuse formulae:\n";
3245 print_uses(dbgs()));
3247 FilterOutUndesirableDedicatedRegisters();
3248 NarrowSearchSpaceUsingHeuristics();
3250 SmallVector<const Formula *, 8> Solution;
3252 assert(Solution.size() == Uses.size() && "Malformed solution!");
3254 // Release memory that is no longer needed.
3260 // Formulae should be legal.
3261 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3262 E = Uses.end(); I != E; ++I) {
3263 const LSRUse &LU = *I;
3264 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3265 JE = LU.Formulae.end(); J != JE; ++J)
3266 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
3267 LU.Kind, LU.AccessTy, TLI) &&
3268 "Illegal formula generated!");
3272 // Now that we've decided what we want, make it so.
3273 ImplementSolution(Solution, P);
3276 void LSRInstance::print_factors_and_types(raw_ostream &OS) const {
3277 if (Factors.empty() && Types.empty()) return;
3279 OS << "LSR has identified the following interesting factors and types: ";
3282 for (SmallSetVector<int64_t, 8>::const_iterator
3283 I = Factors.begin(), E = Factors.end(); I != E; ++I) {
3284 if (!First) OS << ", ";
3289 for (SmallSetVector<const Type *, 4>::const_iterator
3290 I = Types.begin(), E = Types.end(); I != E; ++I) {
3291 if (!First) OS << ", ";
3293 OS << '(' << **I << ')';
3298 void LSRInstance::print_fixups(raw_ostream &OS) const {
3299 OS << "LSR is examining the following fixup sites:\n";
3300 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(),
3301 E = Fixups.end(); I != E; ++I) {
3302 const LSRFixup &LF = *I;
3309 void LSRInstance::print_uses(raw_ostream &OS) const {
3310 OS << "LSR is examining the following uses:\n";
3311 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(),
3312 E = Uses.end(); I != E; ++I) {
3313 const LSRUse &LU = *I;
3317 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
3318 JE = LU.Formulae.end(); J != JE; ++J) {
3326 void LSRInstance::print(raw_ostream &OS) const {
3327 print_factors_and_types(OS);
3332 void LSRInstance::dump() const {
3333 print(errs()); errs() << '\n';
3338 class LoopStrengthReduce : public LoopPass {
3339 /// TLI - Keep a pointer of a TargetLowering to consult for determining
3340 /// transformation profitability.
3341 const TargetLowering *const TLI;
3344 static char ID; // Pass ID, replacement for typeid
3345 explicit LoopStrengthReduce(const TargetLowering *tli = 0);
3348 bool runOnLoop(Loop *L, LPPassManager &LPM);
3349 void getAnalysisUsage(AnalysisUsage &AU) const;
3354 char LoopStrengthReduce::ID = 0;
3355 static RegisterPass<LoopStrengthReduce>
3356 X("loop-reduce", "Loop Strength Reduction");
3358 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
3359 return new LoopStrengthReduce(TLI);
3362 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
3363 : LoopPass(&ID), TLI(tli) {}
3365 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const {
3366 // We split critical edges, so we change the CFG. However, we do update
3367 // many analyses if they are around.
3368 AU.addPreservedID(LoopSimplifyID);
3369 AU.addPreserved("domfrontier");
3371 AU.addRequired<LoopInfo>();
3372 AU.addPreserved<LoopInfo>();
3373 AU.addRequiredID(LoopSimplifyID);
3374 AU.addRequired<DominatorTree>();
3375 AU.addPreserved<DominatorTree>();
3376 AU.addRequired<ScalarEvolution>();
3377 AU.addPreserved<ScalarEvolution>();
3378 AU.addRequired<IVUsers>();
3379 AU.addPreserved<IVUsers>();
3382 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
3383 bool Changed = false;
3385 // Run the main LSR transformation.
3386 Changed |= LSRInstance(TLI, L, this).getChanged();
3388 // At this point, it is worth checking to see if any recurrence PHIs are also
3389 // dead, so that we can remove them as well.
3390 Changed |= DeleteDeadPHIs(L->getHeader());